Re: rtx_unchanging_p vs c++ vtable fields

[kenner at vlsi1 dot ultra dot nyu dot edu (Richard Kenner)]

    I didn't think the field was marked TREE_READONLY, but I don't
    know for sure.  It doesn't get marked readonly at the point of
    creation, but that doesn't mean it doesn't get set elsewhere.

Well that's what readonly_fields_p tests, so if the code in question
is triggered, it *must* have been set.

    I think we should get rid of /u.  It is _way_ too error prone.
    I'll claim that virtually all of the benefit can be gained by
    proper use of alias sets, and if there's anything left oh well.

I think that's right, but it's going to be a bit of work.  I did a large part
of it by putting constants into their own alias set.  That will get a large
part of the benefit for C, but probably not for Ada and I'm not sure about
C++.  Also, if we go that route, we should delete the in-struct and in-scalar,
which historically have have caused even more problems.  In that situation,
I think we should use aliasing at -O1: it's not an expensive optimization.